Author: Acevedo, Stefanie
Title: Harmonic Schemata of Popular Music: An Empirical Investigation of Analytical Patterns and their Mental Representations
Institution: Yale University
Begun: August 2012
Completed: May 2020
A hallmark of music analysis is the identification and categorization of musical patterns into stylistic templates. These templates can be used for evaluating other pieces or comparing music across styles. This analytical process is akin to schematic learning in psychology: repeated experiences develop probabilisitic templates that guide expectations for future events. Statistical learning is the process whereby probabilities of events determine these mental templates, or schemata.
Given an assumption of statistical learning, it is possible to study the probabilities of a musical style in order to model the expectations established from exposure to that corpus of music. These statistical properties are quantified as information content, which is measured as information entropy. Entropy is used in this dissertation to segment common harmonic patterns in a popular music corpus (the McGill Billboard Corpus, a sample of the Top 100 Billboard Hits from 1958-1991). Assuming statistical learning of musical style due to mass proliferation of the songs in this corpus, the resultant harmonic patterns are taken as probabilistic templates, or schemata, that guide harmonic expectations.
The first chapter reviews relevant popular harmony studies and literature on computational musical analysis, including the use of information theory for style analysis. The study of popular harmony relies heavily on comparisons to harmonic syntax of common-practice tonality, namely identifying how and why harmonic templates deviate from cadential phrase models. I argue that many recent approaches are unable to fully explain harmony in popular music due to this biased lens. The rest of the chapter provides a background in computational analyses of language and music, focusing on the use of probabilistic measures for the segmentation of language and applications to music (including previous studies employing the IDyOM model).
Chapters 2 and 3 discuss the main methodologies of this study. Chapter two describes the basic parsing stages of the computational analysis (including n-gram tree building) and provides descriptive statistics of the McGill Billboard Corpus. Based on zeroeth- and first-order harmonic probabilities, the McGill Billboard corpus is compared to other popular and non-popular corpora (namely the Rolling Stone Corpus, Essen Folksong Corpus, and Yale Classical Archive Corpus). Given their limited context and reliance on a specific tonal lens (such as the centering on a tonic), I argue that these types of analyses cannot model harmonic expectation. Chapter three continues the analyses started in chapter 2, implementing entropy as a means to segment harmonic patterns into licks, correlating high uncertainty with the end of a progression. The resultant licks are labeled according to chord quality and interval distance between chord roots, therefore eliminating the need for identification according to a tonal center or mode. Over 1,500 entropy end-bounded licks are identified, but these are culled to the most representative licks (according to the number of songs in which the appear). The licks are organized into 15 families according to intervalic patterns between chords.
The final chapter develops a framework for understanding harmonic behavior beyond traditional functional phrase models. Building on previous popular harmony literature (namely by Allan Moore, Nicole Biamonte, and Philip Tagg), I develop two categories of harmonic behavior: transitional and prolongational functions. These are categories of harmonic progressions, dependent on the distance traversed through tonal space; prolongational patterns equate to zero displacement (begin and end on the same chord), and transitional patterns to non-zero displacement (begin and end on different chords). These categories allow any harmonic pattern to be classified as functional despite tonality or syntax. Families of licks introduced in chapter 3, as well as multiple progressions taken from the literature, are categorized in this system and equated to harmonic schematic templates that guide listening. To support these claims, I analyze a segment from Tina Turner's "What's Love Got to Do with It?", employing the lick families and new functional model to show their applicability despite the musical example's ambiguous tonality (and form).
The dissertation ends by reflecting on the generalizability of the results. Since harmony occurs within multi-variable system that includes form, rhythm, meter, and timbre, the method provided here is limited. In order to truly model learning and expectation of harmony, the model can and should be expanded to include other musical parameters. Further, the mental templates proposed (whether licks or functional categories) provide testable hypotheses for follow-up work on harmonic expectation. Both are avenues for future research that will help develop a better understanding of statistical learning and expectation for popular and other musics.
Keywords: cognition of harmony, popular harmony, harmonic schemata, information entropy, segmentation, popular music, expectation, mental template, computational music theory
Chapter 1 - Background Research in Popular Harmony and Segmentation
Chapter 2 - A Descriptive Statistical Analysis of the McGill Billboard Corpus
Chapter 3 - Harmonic Expectation and Cross-Entropy as a Tool for Chord Sequence Segmentation
Chapter 4 - Progressions as Categories of Harmonic Behavior